# BEVFormer v2: Adapting Modern Image Backbones to Bird's-Eye-View Recognition via Perspective Supervision

**Authors:** Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, Jie Zhou, Jifeng Dai  
**arXiv:** [2211.10439](https://arxiv.org/abs/2211.10439)  
**Date:** 2022-11-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2211.10439v1/

## Key points

- Authors present a novel bird's-eye-view (BEV) detector with perspective supervision
- Proposed detector overcomes limitations of existing BEV detectors tied to specific depth pre-trained backbones
- Introduce perspective space supervision to ease optimization of BEV detectors
- Two-stage BEV detector where proposals from perspective head are fed into bird's-eye-view head for final predictions
- Faster convergence and better compatibility with modern image backbones
- Extensive ablation studies conducted on different forms of supervision and generality of proposed detector
- Achieved new state-of-the-art results on large scale nuScenes dataset
- Innovative approach to adapting modern image backbones to bird's eye view recognition by introducing perspective supervision

## Summary

In their paper titled "BEVFormer v2: Adapting Modern Image Backbones to Bird's-Eye-View Recognition via Perspective Supervision," authors Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu and Jie Zhou and Jifeng Dai present a novel bird's-eye-view (BEV) detector with perspective supervision. The proposed detector aims to overcome the limitations of existing state-of-the-art BEV detectors that are tied to specific depth pre-trained backbones like VoVNet. This limitation hinders the synergy between booming image backbones and BEV detectors. To address this issue, the authors introduce perspective space supervision to ease the optimization of BEV detectors. They propose a two-stage BEV detector where proposals from the perspective head are fed into the bird's-eye-view head for final predictions. This approach allows for faster convergence and better compatibility with modern image backbones. The effectiveness of the model is evaluated through extensive ablation studies focusing on different forms of supervision and the generality of the proposed detector. The authors verify their method using a wide range of traditional and modern image backbones and achieve new state-of-the-art results on the large scale nuScenes dataset. Overall, this paper presents an innovative approach to adapting modern image backbones to bird's eye view recognition by introducing perspective supervision. The proposed method shows promising results and opens up possibilities for further advancements in BEV detection. The code for this research will be released soon.